Yao Maoying, from the tech company Mifeng, said: "The world model needs to be able to predict the next step. Currently, there are two main problems."

On the afternoon of July 17th, during the main forum of the World Artificial Intelligence Conference, with the theme "From Virtual to Reality: How World Models Drive Embodied Intelligence", Yao Maoqing, partner and senior vice president of Zhielement, president of the Embodied Business Department, and chairman and CEO of MiFeng Technology, stated that the essence of a world model is an AI system that can understand the laws of the physical world, with the most important function being able to predict the next state of the world. Although world models have gained popularity this year, most of the work is still focused on visual generation, mainly based on post-training of pre-trained models for video generation." Yao Maoqing pointed out that there are two major problems in current world model training: one is the mismatch of data types, as the distribution of data from internet videos is fundamentally different from the physical world data that robots truly need; and the other is the lack of data scale.